Fish density estimation using unbaited cameras: Accounting for environmental-dependent detectability

摘要

The fast development of camera technologies opens a breakthrough opportunity for animal ecology, particularly at the marine realm where observing wildlife is challenging. These outstanding technological advances are meeting with the impressive capabilities of artificial intelligence for enabling automatic extraction of relevant information from videos and images. Altogether, this may be a unique opportunity for a qualitative jump in marine wildlife assessment but substantial strengthening of the links between theorists, empiricists and engineers is still required. Specifically, a recent theory proposes that animal density can be estimated from (1) the counted animals per frame, (2) the area surveyed by the camera and (3) the probability of detecting an animal that is actually within the area surveyed by the camera. However, a potential drawback for applying this theory to the real world is that environmental dependencies of camera's detection probability may lead to biased estimates of animal density. Therefore, here we propose a sampling protocol and a statistical model of general application for estimating (and accounting for) the environmental factors affecting fish detectability when estimating fish density with cameras. The method implies one calibration sampling with cameras and with the preferred reference method at the same time and place. The relevance of this method is that, once calibrated, it can be used to obtain unbiased estimates of fish density at new sites and moments using only cameras. Thus, fish density could be estimated at the temporal and the spatial scale needed, but with substantially less cost-effort than any other reference methods (e.g., underwater visual censuses). As a proof of concept, we evaluated the dependence of camera's detection probability on habitat complexity (e.g., cavities, rocks, seagrass, etc.) as a proxy for the hiding capability of a small serranid. In that specific case, probability of detection seems to be independent of habitat complexity. However, the sampling protocol and the statistical model provided here open the opportunity to estimate fish density using underwater cameras at wider temporal and/or spatial scales, which will help to better understanding the ultimate drivers of marine fish population dynamics and further development of science-based management.